songbirdini/v33da_pp

V33DA++: Overlapping-Call and Long-Context Companion to V33DA

V33DA++ is the companion release to the V33DA benchmark. It is built from the same BirdPark recordings and the same 10 zebra finches, but covers data outside the strict single-vocalizer attribution clips that define V33DA. V33DA++ is intended for tasks that benefit from overlapping callers or longer temporal context: source separation, vocal activity detection, audio-visual synchronization, and active-speaker detection.

V33DA++ ships three buckets, each as a separate config:

Bucket	Events	Description
overlap	3,784	Reviewer-verified events excluded from V33DA because another bird vocalized in the same window. Each sample carries the full list of overlapping callers verified from on-body accelerometer channels.
padded	33,625	±2 s context windows around V33DA single-vocalizer events. Original event onsets and offsets are preserved inside the longer window, supplying long-range context without altering the underlying call.
overlap_padded	3,784	The overlap bucket with the same ±2 s context, for long-context multi-caller tasks.

V33DA++ is not the benchmark used in the V33DA paper's main results tables. For the single-vocalizer attribution benchmark, use V33DA.

Code: github.com/marisbasha/v33da  ·  Main benchmark: songbirdini/V33DA

Quick start

The parquet shards are the canonical V33DA++ release and contain inline audio + accelerometer bytes for every event — they are self-sufficient. The optional flat audio/, accelerometer/, and clips/ mirrors are convenience exports (see "Release contents" below).

Load via datasets:

from datasets import load_dataset

# Pick one of: "overlap", "padded", "overlap_padded"
ds = load_dataset("songbirdini/v33da_pp", "padded", split="train", streaming=True)
for row in ds.take(1):
    print(row.keys())

Or download a single bucket:

pip install huggingface_hub

# Just the parquet for a bucket (smallest, most useful)
huggingface-cli download songbirdini/v33da_pp --repo-type dataset \
    --include "padded/*.parquet" --local-dir data/v33da_pp

# Full bucket (parquet + media mirrors)
huggingface-cli download songbirdini/v33da_pp --repo-type dataset \
    --include "padded/*" --local-dir data/v33da_pp

Release contents

overlap/
  v33da_pp_overlap-*.parquet              # canonical: includes inline audio bytes
  audio/<experiment>/...                  # convenience WAV mirror (optional)
  accelerometer/<experiment>/...          # convenience WAV mirror (optional)
  clips/<experiment>/<session>/...        # composite multi-view MP4s
  metadata.json
padded/
  v33da_pp_padded-*.parquet
  audio/, accelerometer/, clips/
  metadata.json
overlap_padded/
  v33da_pp_overlap_padded-*.parquet
  audio/, accelerometer/, clips/
  metadata.json

The flat WAV mirrors are best-effort: HuggingFace caps each directory at 10 000 files, so for padded/audio/juvExpBP01 and padded/audio/juvExpBP02 (and the matching accelerometer/ directories) the mirror is truncated. The parquet shards are unaffected and contain every event in full.

Per-bird counts

Experiment	Bird	overlap	padded	overlap_padded
BP01	blue	999	5,735	999
BP01	peach	160	2,277	160
BP01	red	404	3,963	404
BP01	white	521	4,244	521
BP02	blue	915	4,923	915
BP02	peach	292	4,589	292
BP02	red	460	2,065	460
BP05	brown	1	56	1
BP05	purple	20	2,140	20
BP05	yellow	12	3,633	12
Total		3,784	33,625	3,784

Data fields (parquet)

Each parquet row is one vocalization event. Most binary array columns are NumPy np.save() output; load them with np.load(io.BytesIO(row[col])).

Identity & metadata

Field	Type	Description
id	string	Globally unique row identifier: "{experiment}/{video_stem}/{mid_frame}/{bird_name}"
experiment	string	One of juvExpBP01, juvExpBP02, juvExpBP05
date	string	Recording date YYYY-MM-DD
video_stem	string	Source recording stem (e.g. BP_2021-08-05_08-40-49_416952_0460000)
bird_name	string	Internal vocalizer ID (per-experiment, e.g. b14o14_j)
bird_color	string	Vocalizer's backpack color (per-experiment identifier; not global)
vocalizer_idx	int32	Vocalizer's index within this experiment's bird set (0-based)

Event timing (this row's vocalization)

Field	Type	Description
event_onset_sec	float32	Voc onset in seconds, relative to the source recording
event_offset_sec	float32	Voc offset in seconds, relative to the source recording
event_onset_frame	int32	Voc onset in 47.68 fps video frames
event_offset_frame	int32	Voc offset in 47.68 fps video frames
context_pad_sec	float32	Audio/video pad on each side of the event (≈0.01 for overlap, 2.0 for the *_padded buckets)
n_frames	int32	Number of video frames covered by the keypoint window

Audio & accelerometer

Field	Type	Description
audio	bytes	NumPy-serialized (5, T) float32 array — 5 cage microphones at 24,414 Hz, padded by context_pad_sec on each side
audio_path	string	Path to a separately exported multichannel WAV mirror (best-effort, see Release contents)
accelerometer	bytes	NumPy-serialized (N_birds, T) float32 array — per-bird on-body vibration channel at 24,414 Hz, used as the verification oracle
accelerometer_path	string	Path to a separately exported multichannel WAV mirror (best-effort)
video_path	string	Path to a composite multi-view MP4 (top, back, side) — null if not available

3D pose (every visible bird, every frame)

5 keypoints per bird: beak, head, backpack, tail base, tail tip.

Field	Type	Description
keypoints_3d	bytes	NumPy (T_frames, N_birds, 5, 3) float32 — 3D positions triangulated from top + back views, in BirdPark cage coordinates (mm)
keypoints_2d_top	bytes	NumPy (T_frames, N_birds, 5, 2) float32 — 2D keypoints in the top-camera image (pixels)
keypoints_2d_back	bytes	NumPy (T_frames, N_birds, 5, 2) float32 — 2D keypoints in the back-camera image (pixels)

Per-frame indexing: bird i's keypoint k at frame t is kp3d[t, i, k] (and similarly for the 2D arrays). NaN values may appear in rows that were extracted under a lenient filter (see Lenient extraction, below). Bird index → bird color/name mapping is fixed within an experiment (see metadata.json).

Radio telemetry (every visible bird, every frame)

Per-bird FM radio signals from a 4-antenna receiver array (4 antennas: A, B, C, D; M = mean over antennas). Each is a NumPy (N_birds, T_radio) float32 byte blob.

Field	Description
radio_frq	Per-bird FM carrier frequency offset (Hz)
radio_thetaS	Source-side phase used for triangulation
radio_thetaA / B / C / D	Per-antenna received phase (rad)
radio_phiA / B / C / D / M	Per-antenna and mean phase difference (rad)
radio_powRA / B / C / D / M	Per-antenna and mean received power (linear)
radio_powNA / B / C / D / M	Per-antenna and mean noise floor (linear)

Overlap-event metadata (overlap and overlap_padded only)

These rows belong to multi-caller events: the audio window contains vocalizations from ≥2 birds at once. The fields below describe every concurrent vocalization that overlapped with this row's bird in any +/-PAD_SEC sliding window (including the row's own voc), enabling per-caller source separation, per-caller vocal-activity-detection, etc.

There are two views of the same multi-caller event, kept separate for compatibility:

Color-set view (existing fields, unique colors)

Field	Type	Description
overlap_callers	list[string]	Backpack colors of all concurrent callers, deduplicated and sorted alphabetically (e.g. ['blue', 'white'])
overlap_count	int32	len(overlap_callers) — number of unique caller-birds in the event

Vocalization-list view (new fields, one entry per voc)

These six lists are parallel (same length, same index order — sorted by bird_color, ties broken by onset). Length is overlap_voc_count, which can be larger than overlap_count if the same bird vocalized more than once in the active window or if the row participated in two adjacent overlap windows that share a partner.

Field	Type	Description
overlap_caller_bird_names	list[string]	Internal bird names (e.g. 'b14o14_j') — one per voc
overlap_caller_bird_colors	list[string]	Backpack colors (may contain duplicates), same order as bird_names
overlap_caller_onset_sec	list[float32]	Each voc's onset in seconds (same time frame as event_onset_sec)
overlap_caller_offset_sec	list[float32]	Each voc's offset in seconds
overlap_caller_onset_frame	list[int32]	Each voc's onset in 47.68 fps video frames
overlap_caller_offset_frame	list[int32]	Each voc's offset in 47.68 fps video frames
overlap_voc_count	int32	len(overlap_caller_bird_names)

Tip. To find when caller 'white' is vocalizing inside this event, do i = overlap_caller_bird_colors.index('white'); on, off = overlap_caller_onset_sec[i], overlap_caller_offset_sec[i]. If 'white' appears multiple times (multiple vocs from the same bird), iterate through all matches.

Lenient extraction

A small fraction of rows (~6% of overlap and overlap_padded) are extracted under a lenient policy: the per-frame motion-displacement filter is bypassed and any NaN keypoints are zero-filled. These rows have valid audio, radio, accelerometer, and event-timing fields but possibly noisier 3D pose. They are included so the bucket size matches the paper's event count.

You can recognise them by checking for any zeros (or large displacements) in keypoints_3d. Strictly extracted rows (under V33DA's normal filter) have all-finite keypoints with frame-to-frame displacement ≤ 40 px.

Per-bucket metadata

Each bucket ships a metadata.json listing exact row counts, per-bird counts, schema, filter settings, frame rate, and pad-frame counts.

Note on bird identity. Backpack color codes are experiment-local identifiers, not global bird IDs. A color reused across experiments (e.g. red in BP01 and BP02) refers to two different physical individuals; all 10 released birds are distinct.

Suggested tasks

• Source separation — given multichannel audio with 2+ concurrent callers, recover per-bird streams (overlap bucket; overlap_callers provides the speaker set).
• Voice activity detection — given padded multichannel audio, predict per-frame call/non-call (padded and overlap_padded buckets; event_onset/offset_sec provide reference boundaries).
• Active-speaker detection / audio-visual sync — given multi-view video and multichannel audio in a long window, predict who is calling when (overlap_padded bucket).

License

CC-BY-4.0. External tools used during preprocessing (e.g., Whisper / WhisperSeg, SAM2, ByteTrack, YOLOX-Pose) retain their original licenses.

Citation

If you use V33DA or V33DA++, please cite the V33DA paper:

@unpublished{basha2026v33da,
  title  = {Who Called? V33DA: A Physically Verified Multimodal Benchmark for Vocal Attribution in Zebra Finch Groups},
  author = {Basha, Maris and Wang, Yuhang and Chen, Xiaoran and Cheng, Longbiao and Yapura, Luca and Zai, Anja T. and Salzmann, Mathieu and Hahnloser, Richard},
  year   = {2026},
  note   = {Under review},
}